Here's what the feature does, how to turn it on, how it can go wrong, and how to pilot it without letting an AI model click the wrong button in production.
What the roadmap entry says
Roadmap item 579906, "Power Automate: Self-healing," calls the capability an agentic feature that makes desktop flows more resilient by recovering from UI and web automation failures. When a saved selector can no longer find its target element, Power Automate for desktop uses AI to identify the element and generate an alternative selector.
The entry describes two phases:
- Preview: You opt in per action, and the flow automatically tries to recover from a selector error.
- General availability: Adds a flow-level opt-in that only logs the working alternative selector when an error occurs. It doesn't apply the fix at runtime.
The item is marked In development. It lists April 2026 for preview and March 2027 for GA, on the Desktop platform in the Worldwide (Standard Multi-Tenant) cloud. Treat March 2027 as an estimate. Microsoft says roadmap information is subject to change, and features can be postponed or cancelled.
Section summary: The preview offers automatic repair per action. The GA phase is set to add a flow-wide "suggest, don't touch" mode, which is the more interesting addition for governance-minded teams.
The preview is already out, and earlier than the roadmap says
The preview has been shipping for months, and its start predates the roadmap's April 2026 date. Microsoft's release notes for Power Automate for desktop Build 2603 say self-healing was introduced in preview as an AI-powered, agentic feature in Power Automate for desktop, designed to make UI and browser automations more resilient when a captured element can't be found during runtime. That build was installer version 2.66.149.26061 (Microsoft Store version 11.2603.149.0), and the feature was set to gradually roll out in various regions during March.
Microsoft's 2025 release wave 2 plan lists public preview for the "Self-healing agent for UI/web automation in desktop flows" as enabled for admins, makers, marketers, or analysts, automatically, on Mar 31, 2026, with no GA date. So the release plan doesn't independently confirm the roadmap's March 2027 date.
The idea isn't new either. A 2024 release wave 1 item, summarised by the azurecurve blog, described a feature to self-heal UI and browser automation actions at execution with AI, with a planned public preview in Jun 2024 and general availability in Sep 2024. That version said the repaired selector is used in future executions. Those 2024 dates didn't result in the feature documented today. The current version arrived in 2026 with a different design, so the March 2027 estimate deserves some caution too.
How self-healing works at runtime
Microsoft Learn says the preview attempts to identify the most likely correct element at runtime and continue the flow run, instead of stopping the flow when an element can't be found. It targets two common failure modes: an app or page changes an element's underlying identifier, or a control looks the same to a person but no longer matches its saved selector.
Self-healing is one step in a fixed order of error handling, not a replacement for it:
Retry policy → Self-healing (preview) → 'Set variable / Run subflow' rules → Continue flow run / Throw error
In practice:
- Your configured retry policy runs first.
- If the retries fail, self-healing captures context: a screenshot of the missing element, the title of its parent window, and a full desktop image. It uses these to find the likely target and capture a new selector.
- If the repair works, the flow continues. If not, your normal "Set variable / Run subflow" rules and continue-or-throw logic take over.
According to the October 2026 version of the Microsoft Learn page, two models currently work together to power the feature: GPT-4.1 mini and Claude Sonnet 4.5. That could change, so don't hard-code it into a compliance document.
Self-healing works only at runtime. Self-healing is a runtime-only capability. It applies to console runs and cloud-triggered runs, attended and unattended, and does nothing during designer runs. If you test it by clicking Run in the designer and nothing happens, that's expected behavior.
Scope: what it covers and what it doesn't
The preview has narrow limits:
| Covered | Not covered |
|---|---|
| UI and browser automation actions that act on a single element (clicking a button, selecting a checkbox) | Window- and screen-handling actions |
| "Element not found" errors on primary and secondary screens | Multi-element actions such as "Drag and drop UI element in window" |
| "Window not found" errors, only when the window is on the main screen | Designer runs |
| Console and cloud runs, attended and unattended | Government clouds (GCC, GCC High, DoD) and sovereign clouds |
Microsoft's own documentation isn't fully consistent here. The general "Handle errors in desktop flows" page says self-healing is currently available only for specific UI and browser automation actions that interact with a single UI element, and applies only to Element not found errors. The dedicated self-healing page, updated October 28, 2026, adds "Window not found" errors with the main-screen restriction. Rely on the dedicated page, but don't count on window-level recovery until you've tested it.
Section summary: This is a targeted fix for single-element selector problems. It doesn't make every desktop flow failure-proof.
Prerequisites and licensing
- Account type: Self-healing in Power Automate for desktop is currently available for organization premium accounts at no additional cost. The current page adds that usage may be subject to service limits or changes.
- Cloud: It isn't currently available in government clouds (GCC, GCC High, DoD) or sovereign clouds.
- Client version: You need Power Automate for desktop version 2.66 or higher. That lines up with Build 2603's 2.66.149 installer.
Admin setup: four switches
The admin side is where pilots usually stall. Self-healing uses generative models, including an Anthropic model, so all of these settings must be on:
- Microsoft 365 admin center: Go to Copilot > Settings > AI providers operating as Microsoft subprocessors and enable Anthropic as a subprocessor. Microsoft enables Anthropic models by default for most commercial-cloud customers, but not in the EU/EFTA or the UK. Tenants there need to opt in on purpose.
- Power Platform admin center: Go to Copilot > Settings. In the Power Platform section, expand External models and enable Anthropic at the environment or environment-group level.
- Environment settings: Under Manage > Environments, open the environment, select Edit on the Generative AI features card, review the terms, and check Move data across regions. This checkbox doesn't appear and isn't needed for US cloud regions.
- Copilot in Power Automate: In Copilot > Settings, under the Power Automate section, select Copilot in Power Automate, then enable the On check box in the displayed pane.
Step 3 deserves attention from privacy and compliance teams. Self-healing captures screenshots and full-desktop images when it runs. If your automations work with screens showing patient records, payroll, or customer PII, decide whether those images can cross regions and go to a third-party subprocessor before anyone enables this. That's my analysis, based on the data Microsoft says the feature collects, not a Microsoft compliance statement.
Maker setup: turn it on per action
In the flow designer:
- Select a supported UI or browser automation action. If self-healing is available for it, an indicator appears when you hover over the action.
- Open the action's Error handling settings.
- Turn on Self-healing (Preview).
When it's on, the action can retry with AI if the element isn't found at runtime.
Common problems
- Toggle missing? Check that you're on 2.66 or later and that the action works on a single element. Then check all four admin settings above. A missing toggle usually points to an admin setting that's off.
- Nothing happening in tests? You're probably testing in the designer. Run the flow from the console or trigger it from a cloud flow.
- Still failing? Your retry policy runs first. If self-healing can't find a confident match, your standard error rules apply.
- EU/UK tenant blocked? Anthropic isn't on by default for you. Check step 1.
Monitoring: check what the AI chose
Microsoft states the risk directly: because self-healing uses generative AI, it might occasionally identify an unintended UI element. That's acceptable for a "Next" button. It's a serious problem if the alternative to "Cancel order" turns out to be "Approve order."
Visibility has improved but is still limited:
- The Power Automate portal doesn't currently surface self-healing details in the action-level logs of the flow runs where self-healing was invoked. In the desktop flow run page, you can find a summary of the self-healing activity that corresponds to that specific run, including details from up to three successful self-healing attempts.
- With a registry setting Microsoft documents, local action logs on the machine that ran the flow include a
repairWithAIInfoproperty. It shows whether a repair was attempted, whether it worked, and which AI-generated selector was used. - Build 2607 (installer 2.70.184.26183) added broader tolerance during selector validation, so more automatic repairs are applied. It also added portal reporting of whether self-healing was on for each run and how many repairs succeeded.
Note the trade-off in that last item. Broader tolerance means more successful repairs, but more permissive matching also means each repair needs a closer look. It's worth checking those repair counts regularly.
Don't confuse it with AI-assisted repair
Build 2607 also introduced a separate preview called AI-assisted UI automation repair. It's a maker-led debugging tool in the flow designer. It finds broken selectors, proposes AI-generated replacements after you recapture the element, and lets you test, accept, or skip each one. Accepted selectors go into the control repository. Self-healing works unattended at runtime. AI-assisted repair works with a person at design time.
The GA "log only" mode fits between these two. A runtime fix that's recorded but not applied gives makers a tested candidate selector to review in the designer, without the AI acting on its own in production.
Analysis: who should pilot this
Self-healing addresses the biggest weakness of UI-based automation. Selectors break whenever the UI changes, and automation teams spend real time on that maintenance. Moving from a hard stop to an AI-assisted fallback is a meaningful change, especially for unattended flows that fail overnight.
As the consultancy FlowDevs put it, the capability is still in preview, applies to specific failure types, and should not be treated like permission to ignore process design, testing, or oversight.
A sensible rollout:
- Start with low-risk, read-heavy actions: navigation, opening menus, ticking filters.
- Leave it off for irreversible actions like submit, approve, delete, and payment buttons, at least until the GA log-only mode lets you review fixes before they're applied.
- Add validation checks after repaired steps, such as confirming that the expected window title or field value is present.
- Review repair data weekly and update the selector permanently through proper maintenance. A frequently repaired selector is technical debt.
- Involve privacy teams early about screenshots, subprocessors, and cross-region data movement.
The roadmap's March 2027 GA date is an estimate and could change. Either way, desktop flows can now attempt recovery instead of stopping at the first changed button, and enterprises adopting it need to put checks in place for the cases where the AI picks the wrong element.
References
- Power Automate: Self-healing Microsoft 365 Roadmap · 2026-11-17T18:36:41.967573Z
- New Functionality In Microsoft Power Automate 2024 Wave 1: Self-heal UI and Browser Automation Actions at Execution with AI azurecurve.co.uk
- Self-healing (preview) - Power Automate learn.microsoft.com